Learning a CNN with Class Specific Backpropagation Computer

Learning a CNN with Class. Specific Backpropagation Computer Vision Lab. Hyeonseob Nam

Motivation • Learning an object detector is often disturbed by noisy regions in the training data.

Goal • Build a robust object detector by learning a CNN with class-specific regions.

Related Works Maxime Oquab et al. , Weakly supervised object recognition with convolutional neural networks, NIPS 2014

Related Works Deepak Pathak et al. , Fully Convolutional Multi-Class Multiple Instance Learning, ar. Xiv 2015

Related Works Karen Simonya et al. , Deep inside convolutional networks: Visualising image classification models and saliency maps, ar. Xiv 2014

Forward-propagation in a CNN

Back-propagation in a CNN

Back-propagation in a CNN

Back-propagation in a CNN

Class-specific Back-propagation in a CNN

Class-specific Back-propagation in a CNN

Class-specific Back-propagation in a CNN

Class-specific Back-propagation in a CNN

Class-specific Back-propagation in a CNN input image pool 3 pool 4 pool 5 score

Outline Pretraining in Imagenet Finetuning with class-specific backpropagation rcnn-based detection

Thank you